Tag: agentic AI safety
Browse our exclusive articles!
Secure LLM Instructions: Automated Framework Against Encoding Attacks
Discover an automated framework to evaluate and harden LLM system instructions, protecting sensitive data from encoding attacks without costly retraining.
Addressing Vulnerabilities in Aligned AI Systems
Explore key vulnerabilities in aligned AI systems and innovative strategies like ACDC and LAT to enhance AI safety and reliability.
Robust Multimodal Safety with CASA Conditional Decoding
Enhance multimodal model safety using CASA's conditional decoding and safety attention to reduce harmful query risks without extra classifiers.
AI Safety Gates: Why Classifier-Based Methods Fail
Study reveals classifier-based AI safety gates fail in self-improving systems; Lipschitz ball verification offers a safer alternative.
How Language Models Process Ethical Instructions: Key Insights
Explore how top language models process ethical instructions, revealing distinct types of ethical reasoning and the impact of instruction formats.